Unboxing the black boxes (Deprecated version)

September, 2018 @leonardo_noleto
Leonardo Noleto
Senior Data Scientist
Unboxing
the Black Boxes

Confidential - All rights reserved, any reproduction or distribution of the content is prohibited without Bleckwen’s written permission
Leonardo Noleto
Ϝ(𝑐𝑜𝑚𝑝𝑢𝑡𝑒𝑟 𝑠𝑐𝑖𝑒𝑛𝑡𝑖𝑠𝑡) → 𝑑𝑎𝑡𝑎 𝑠𝑐𝑖𝑒𝑛𝑡𝑖𝑠𝑡
Former software engineer.
Senior data scientist @ Bleckwen
Previously worked for start-ups and large
companies (OVH and KPMG)
co-founder and former organizer of
Toulouse Data Science Meetup
About Me
2
@leonardo_noleto
linkedin.com/in/noleto

Machine learning models are everywhere
Machine learning is at the core of many recent advances in
science and technology

Booking an
appointment for you

Beating the Go Master

Carrying goods

Defeating Breast Cancer

Speaking Chinese

And even painting

Confidential - All rights reserved, any reproduction or distribution of the content is prohibited without Bleckwen’s written permission 10
But sometimes things go wrong.…
A major flaw in Google's algorithm
allegedly tagged two black people's faces
with the word 'gorillas'
Source: http://www.businessinsider.fr/us/google-tags-black-people-as-gorillas-2015-7

But sometimes things go very, very badly wrong
There’s software used across the
country (US) to predict future criminals.
And it’s biased against blacks.
Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing

• With the growing impact of ML algorithms on
society, it is no longer acceptable to trust the
model without an answer to the question: why?
why did the model make a precise decision?
• Data science professionals used to focus on
metrics like accuracy and model generalization
• We need to make sure that models we put in prod
do not hurt people
Machine learning is becoming ubiquitous in our lives
xkcd on Machine Learning

Image from Drive.ai, a self-driving car service for public use in Frisco, Texas
Why we should care about Interpretability?

“Interpretability is the degree to which a human
can understand the cause of a decision.”
Miller, Tim. 2017. “Explanation in Artificial Intelligence: Insights from the Social Sciences.”

Interpretability, why it is important?
STRENGTHEN TRUST AND
TRANSPARENCY
A model may discriminate against certain
populations or make decisions based on
fallacious correlations. Without
understanding the captured factors, we will
have no guarantee that decisions will be fair.
SATISFY REGULATORY
REQUIREMENTS
The GDPR determines how the personal data
of European citizens can be used and
analyzed. Interpretability is a key enabler for
auditing machine learning models.
EXPLAIN DECISIONS
A machine learning model with better
interpretability allows humans to
establish the diagnosis and understand
what happened.
IMPROVE MODELS
For data scientists, interpretability also
ensures that the model is good for the
right reasons and wrong for the right
reasons as well. In addition, this offers
new possibilities for feature engineering
and debugging the model.

Image source: https://blog.bigml.com/2018/05/01/prediction-explanation-adding-transparency-to-machine-learning/amp/
The trade-off Accuracy vs. Interpretability
No data scientist wants to give up on accuracy…

Explaining the taxonomy of Interpretability
globallocal
Model-agnostic
Model-specific
scope
approach

• In interpretability research there are
two scopes that serve different
purposes, they allow us to understand
the model on two different levels:
• Global: it is often important to understand
the model as a whole and then zoom in on
a specific case (or group of cases). The
overall explanation provides an overview of
the most influential variables in the model
based on the input data and the variable to
be predicted.
• Local: local explanations identify the
specific variables that have contributed to
an individual decision. Note that the most
important variables in global explanation
do not necessarily correspond to the most
important variables in a local prediction
Explaining the taxonomy of Interpretability: Scope
Summarizing Global and Local Interpretation (Source: DataScience.com)

• There are two ways to try and achieve
interpretability:
• Model-agnostic: this approach treats the
model like a black box, meaning that it only
considers the relation between inputs and
outputs
• Model-specific: this approach aims to use
the internal structure of the model itself in
order to explain it (trees, neurons, linear
coefficients...)
Explaining the taxonomy of Interpretability: Approaches

• The first intuition when seeking explanations is using only what we call white box
models (models that are intuitively interpretable) like linear model, decision
trees and rule list.
• The second intuition is using the internal structure of a specific model to extract
reasons behind the decisions. We will go into one example of model specific
approach: one that uses the tree structure to explain tree based models
• As a last resort, we can adopt a model agnostic approach which allows more
liberty in using complex models hopefully without compromising interpretability
How do we enable model interpretability?

Using only white box models

• Probably the most known class of "white box"
model
• Linear models learn a linear and monotonic
functions between features and target
• Linearity makes the model interpretable, we can
deduce the contribution of a feature by looking at
its coefficient in the model
White-box model: Linear Models
monotonically
increasing function.
monotonically
decreasing function.
A function that is not
monotonic
Global
and Local
White
Box

• Pros
• By looking at the coefficients we can get a global idea of what the model is doing
• For one instance we can see how much each feature actually contributes to the final
result
• When interpreting models prefer Ridge instead of Lasso
• Provides regulator-approvable models (i.e.: Equal Credit Opportunity Act - ECOA)
• Cons
• As its name implies, Linear models can only represent linear relationships
• So much unrealistic assumptions for real data (Linearity, normality, homoscedasticity,
independence, fixed features, and absence of multicollinearity)
• Linear models are not that easy to interpret when variables are correlated
• Coefficients cannot easily be compared if input are not standardized
White-box model: Linear Models

• Decision trees often mimic the human
level thinking so it is simple to understand
the data and visualize the logic of
different paths leading to the final
decision
• Interpretation can be reached by reading
the tree as a bunch of nested if-else
statements
White-box model: Decision Trees
Global and
enable Local
White
Box

• Pros
• Can be displayed graphically
• Can be specified as a series of rules, and more closely approximate human decision-making
than other models
• Can automatically learn feature interactions regardless of monotonic transformations
• Tends to ignore irrelevant features
• Cons
• Performance is (generally) not competitive with the best supervised learning methods
• Lack of smoothness
• Can easily overfit the training data (tuning is required)
• Small variations in the data can result in a completely different tree (high variance)
• Suffers with highly unbalanced data sets
White-box model: Decision Trees

• GAMs provides a useful extensions of linear
models, making them more flexible while still
retaining much of their interpretability
• It consists of multiple smoothing functions
(splines)
• Relationships between the individual predictors and
the dependent variable follow smooth patterns that
can be linear or nonlinear
• We can estimate these smooth relationships
simultaneously and then predict g(E(Y))) by simply
adding them up
White-box model: Generalized Additive Models
Image source: https://multithreaded.stitchfix.com/blog/2015/07/30/gam/
Global and
Local
White
Box

• Pros
• Interpretability advantages of Linear Models
• Relationships between independent and dependent variable are not assumed to be linear
• Predictor functions are automatically derived during model estimation
• Regularization of predictor functions helps avoid overfitting
• GAMs are competitive with popular learning techniques like SVM and Random Forest
• Cons
• Can be computationally expensive for large data sets
• One may be tempted to make the model overly complex (with many degrees of freedom) to
get more accuracy (less interpretability and possible over fitting)
• Interpretation is not so straightforward as linear models (more involving)
White-box model: GAMs

Model specific interpretability

• TreeInterpreter is a simple idea that allows us to
have local interpretability for tree based models
(Decision trees, Random Forests, ExtraTrees,
XGBoost…)
• Principle: feature weights are calculated by
following decision paths in trees of an
ensemble. Each node of the tree has an output
score, and contribution of a feature on the
decision path is how much the score changes
from parent to child
• Every prediction can be trivially presented as a
sum of feature contributions, showing how the
features lead to a particular prediction
Model Specific: TreeInterpreter
Local
only
Tree
based
models

Model Specific: Example TreeInterpreter
Let’s take the Boston housing price data set and build a regression decision tree to
predict housing in suburbs of Boston

Model Specific: Example TreeInterpreter
RM LSTAT NOX DIST
3.1 4.5 0.54 1.2
RM: average number of rooms among homes in the neighborhood.
LSTAT: percentage of homeowners in the neighborhood considered "lower class“
NOX: the air quality
DIST: distance from the city center
Target is Median value of owner-occupied homes in $1000s
Prediction: 23.03 ≈ 22.60 (trainset mean) - 2.64(loss from RM) + 3.51(gain from LSTAT) - 0.44(loss from DIS)
Loss from RM
19.96 – 22.60 = -2.64
Loss from DIS
23.03 – 23.47 = -0.44
Gain from LSTAT
23.47 – 19.96 = +3.51
Let’s take the Boston housing price data set and build a regression decision tree to
predict housing in suburbs of Boston

• Pros
• Simple and intuitive
• Enable tree based model local explanation (i.e. why a particular prediction is made)
• Enable tree based model debugging
• Cons
• Original Python package limited to scikit-learn API (use ELI5 for XGBoost)
• Heuristic method (maths behind it is not well founded)
• Biased towards lower splits in the tree (as trees get deeper, this bias only grows)
Model Specific: TreeInterpreter

Model agnostic interpretability

Model agnostic: LIME
• Local Interpretable Model-agnostic Explanations:
explain single predictions of any classifier or
regressor, by approximating it locally with an
interpretable model
• LIME is based on two simple ideas: perturbation
and local surrogate model
Tabular data
Image data
Text data
Local
only
Any
model

Model agnostic: LIME - Intuition
Complex
Model
Perturbation: LIME takes a prediction you want to explain ( ) and systematically perturbs its inputs. This
creates perturbed data in a neighborhood of this data point. Perturbed data become a new labelled training
data set (labels come from the complex model)

Model agnostic: LIME - Intuition
Local surrogate model : LIME then fits an interpretable model (linear model) to describe the relationships
between the (perturbed) inputs and outputs
Local surrogate model

Model agnostic: LIME – How it works
Peek a
sample to
explain
Discretize
numerical
features
(optional)
Create
perturbed
samples
Label
perturbed
data
Feature
selection
Fit a weighted
regression
Extract coefficients
as feature
importance

• Pros
• Quite intuitive (based on surrogate models)
• Creates selective explanations, which humans prefer
• Provides explanation for tabular data, text and images
• Cons
• The library evolved from the paper
• Synthetic data set generation does not reflect original data set distribution
• The explainer can be very unstable
• Tuning parameters is not always clear
• How big should the neighbourhood be?
• Changing numerical discretization method can get to the opposite explanation
• Works poorly for severe class imbalance
Model agnostic: LIME

• Most of the methods that we explored so far are very approximate
• TreeInterpreter for example is based on an intuition and has no formal proof or
justification
• LIME has some theory behind it but makes a lot of assumptions and
approximations, this makes the explanations unstable and unreliable
• We need to go from heuristics and approximations to a more well founded
theory
Model agnostic: Towards a more rigorous approach

• SHAP (SHapley Additive exPlanations) is a unified approach of techniques like
TreeInterpreter, LIME and others to explain the output of any machine learning
model
• SHAP connects coalitional game theory with local explanations so that
explanations satisfies the three following requirements that are desirable in the
task of interpretability: Local accuracy, Missingness and Consistency
Model agnostic: Enters SHAP

• Definition: Coalitional Game theory,
also referred to as cooperative game
theory, is the type of games where
players need to form coalitions
(cooperate) in order to maximize the
general gain
• Let’s take a example of a soccer
championship:
• 11 Players, 1 Coalition, 1 Goal
• They won the tournament
• They also won a huge amount of
money
• How do we distribute the cash prize
fairly among the players?
Model agnostic: SHAP – Introducing coalition game theory

Model agnostic: SHAP – Introducing coalition game theory
« Members should receive
payments proportional to
their marginal contribution »
Lloyd Shapley 1923-2016

Model agnostic: SHAP – Shapely value properties
• Efficiency: The attributions for each
player sum up to the value
• Symmetry: If two players have the
same contribution they have the same
payoff
• Dummy: No value, no payoff
• Additivity: For the sum of two games
the attribution is the sum of
attributions from the two games

Model agnostic: SHAP – The Shapley value theorem
For the coalitional game described by < 𝑁, 𝜈 > there exists a unique attribution 𝜙 that satisfies the
four fairness axioms seen later, it is the Shapley value:
𝜙𝑖 𝜈 =
𝑆⊆𝑁− 𝑖
𝑆 ! 𝑁 − 𝑆 − 1 !
𝑁 !
(𝜈 𝑆 ∪ 𝑖 − 𝜈 𝑆 )
𝜙: 𝑡ℎ𝑒 𝑎𝑡𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛 𝑓𝑜𝑟 𝑝𝑙𝑎𝑦𝑒𝑟 𝑖,
𝜈: 𝑓𝑢𝑛𝑐𝑡𝑖𝑜𝑛 𝑡ℎ𝑎𝑡 𝑟𝑒𝑝𝑟𝑒𝑠𝑒𝑛𝑡𝑠 𝑡ℎ𝑒 ’𝑤𝑜𝑟𝑡ℎ’ 𝑜𝑓 𝑎 𝑐𝑜𝑎𝑙𝑖𝑡𝑖𝑜𝑛,
𝑖: 𝑡ℎ𝑒 𝑝𝑙𝑎𝑦𝑒𝑟 𝑜𝑓 𝑖𝑛𝑡𝑒𝑟𝑒𝑠𝑡
𝑁: 𝐹𝑖𝑛𝑖𝑡𝑒 𝑠𝑒𝑡 𝑜𝑓 𝑎𝑙𝑙 𝑝𝑙𝑎𝑦𝑒𝑟𝑠, 𝑎𝑙𝑠𝑜 𝑐𝑎𝑙𝑙𝑒𝑑 𝑇ℎ𝑒 𝑔𝑟𝑎𝑛𝑑 𝑐𝑜𝑎𝑙𝑖𝑡𝑖𝑜𝑛,
𝑆: 𝑆𝑢𝑏𝑠𝑒𝑡 𝑜𝑓 𝑝𝑙𝑎𝑦𝑒𝑟𝑠, 𝑎 𝑐𝑜𝑎𝑙𝑖𝑡𝑖𝑜𝑛

• From coalitional game theory to model interpretability:
Model agnostic: SHAP – Connecting to Machine Learning
Players
Game
Gain attribution
Features
Model
Feature attribution

• From coalitional game theory to model interpretability:
• Local accuracy: 𝑓 𝑥 = 𝑔(𝑥′) meaning that the model and the explanation model have the
same value for the instance that we want to explain
• Missingness: 𝑥′𝑖 = 0 ⇒ 𝜙𝑖 = 0, meaning that if a value is missing, there’s no importance
attribution
• Consistency: if the value added by a feature is bigger in a modified model, the corresponding
feature attribution is larger. This gives explanations a certain consistency
• All three properties can be equated to the four properties previously seen in the
coalitional game theory. Equivalence can be demonstrated.
Model agnostic: SHAP – Connecting to Machine Learning

• Current feature attribution methods for tree based models are inconsistent
(XGBoost/RF feature importances, Gain split, etc)
• TreeInterpreter needs to be modified to get the right values, the values that
examine the outcomes for every possible subset and match the properties of
Shapely value
• TreeSHAP values are optimal but challenging to compute
• The author of SHAP proposes a novel methods that reduces the computational
cost that is exponential in theory to a polynomial cost
• Detailed computation is beyond the scope of this talk, mathematical explanation
can be found in the author paper: Consistent feature attribution for tree
ensembles
Model agnostic: TreeSHAP
Global
and Local
Tree
based
models

• Inspired by LIME but using the right parameters to get Shapley values
• In the original LIME there are hyperparameters to adjust the vicinity of the
sample to explain, the loss function and the regularization term
• The original LIME choices for these parameters are made heuristically; in general,
it does not meet the properties of local accuracy and consistency
• SHAP’s author propose a new method to find these parameters that are proven
to satisfy Shapely value
• The proof can be found in the author NIPS paper: A Unified Approach to
Interpreting Model Predictions
Model agnostic: KernelSHAP
Global
and Local
Any
model

• LIME helps to debug text classifications and
explain part of the text
• SHAP, LIME and DeepLift provides a method
for decomposing the output prediction of a
neural networks for image classification
Model agnostic: Text and Image Interpretability
ELI5 TextExplainer based on LIME algorithm:
debugging black-box text classifiers
LIME explaining a Bernese mountain dog
labeled from Inception v3 model

• So far, we have seen essentially explainers
that outputs reasons as “feature
contribution”
• There are many other classes of explainers:
• Visual (Partial Dependency Plot and
Individual Conditional Expectations - ICE)
• Rules explanations (Bayesian Rule List,
Anchors, Skope Rules, etc)
• Case reasoning (Influence Funtions)
• Contrastive Explanations (DeepLift,
DeepExplainer, etc)
Model Interpretability: the other families of explainers
Using partial dependence to understand the relationship between
a variable and a model's predictions (source: Skater package)
The plot above explains ten outputs (digits 0-9) for four different images.
Red pixels increase the model's output while blue pixels decrease the
output (source SHAP Deep Explainer)

• Transparency: Make available externally the
design of algorithms, the data collection
procedures and all data treatment
• Fairness: Ensure that algorithmic decisions
do not create discriminatory or unjust
impacts when comparing across different
demographics (e.g. race, sex, etc).
• Accountability : Enable interested third
parties to probe, understand, and review
the behavior of the algorithm through
disclosure of information that enables
monitoring, checking, or criticism
• Privacy: Ensuring that sensitive information
in the data is protected
Beyond Interpretability
Image source: https://www.wired.com/2016/10/understanding-artificial-intelligence-decisions/

• Trade-off:
• Approximate models with straightforward and stable explanations
• Accurate models with approximate explanations
• Evaluating the interpretability quality
• Use simulated data for testing explanations and different packages (methods)
• See O’Reilly article: Testing machine learning interpretability techniques
• Rethink feature engineering
• Is the feature intelligible? How long does it take to understand the explanation?
• It is actionable?
• Is there a monotonicity constraint?
• Model sparsity: How many features are being used by the explanation?
• Computational Speed: agnostic techniques are time-consuming
Lessons learned after 6 months of interpretability

How we are using interpretability techniques
Our clients are fighting against
financial fraud with the power of
Machine Learning and the Behavioral
Profiling while keeping decisions
interpretable

We combine best-in-class Machine Learning models with
interpretation technologies to get the best from a collective
artificial and human intelligence

Collaboration between
Human Being and Artificial Intelligence will
secure the world.

DEEP TRUTH REVEALED
Anti-fraud solution for banks.
Thank you.

• Interpretable Machine Learning: A Guide for Making Black Box Models
Explainable by Christoph Molnar
• O’Reilly Red book: An Introduction to Machine Learning Interpretability
• GAMs: Intelligible Machine Learning Models for HealthCare
• Interpretable ML Symposium: http://interpretable.ml/
• Awesome Interpretable Machine Learning (Github repository)
• FAT/ML Fairness, Accountability, and Transparency in Machine Learning
https://www.fatml.org/
Additional references

BLECKWEN
6 rue Dewoitine
Immeuble Rubis
78140 Vélizy – France
www.bleckwen.ai
Let's stay in touch
Leonardo Noleto
Senior Data Scientist
leonardo.noleto@bleckwen.ai

Unboxing the black boxes (Deprecated version)

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Unboxing the black boxes (Deprecated version)

Similar to Unboxing the black boxes (Deprecated version) (20)

Recently uploaded

Recently uploaded (20)

Unboxing the black boxes (Deprecated version)

Editor's Notes